You are on page 1of 5

Generalized Dose-Response Curves and The Origin of Thin-Tails

(Studies in (ANTI)FRAGILITY) Nassim N. Taleb
The literature of heavy tails starts with a random walk and finds mechanisms that lead to fat tails under aggregation. We follow the inverse route and show how starting with fat tails we get to thin-tails from the probability distribution of the response to a random variable. We introduce a general dose-response curve show how the left amd right-boundedness of the reponse in natural things leads to thin-tails, even when the “underlying” variable of the exposure is fat-tailed.
Very Preliminary Version, July 2013. Comments welcome.

The Origin of Thin Tails.
We have emprisoned the “statistical generator” of things on our planet into the random walk theory: the sum of i.i.d. variables eventually leads to a Gaussian, which is an appealing theory. Or, actually, even worse: at the origin lies a simpler Bernouilli binary generator with variations limited to the set {0,1}, normalized and scaled, under summation. Bernouilli, De Moivre, Galton, Bachelier: all used the mechanism, as illustrated by the Quincunx in which the binomial leads to the Gaussian. This has traditionally been the “generator” mechanism behind everything, from martingales to simple convergence theorems. Every standard textbook teaches the “naturalness” of the thus-obtained Gaussian. In that sense, powerlaws are pathologies. Traditionally, researchers have tried to explain fat tailed distributions using the canonical random walk generator, but twinging it thanks to a series of mechanisms that start with an aggregation of random variables that does not lead to the central limit theorem, owing to lack of independence and the magnification of moves through some mechanism of contagion: preferential attachment, comparative advantage, or, alternatively, rescaling, and similar mechanisms. But the random walk theory fails to accommodate some obvious phenomena. First, many things move by jumps and discontinuities that cannot come from the random walk and the conventional Brownian motion, a theory that proved to be sticky (Mandelbrot, 1997). Second, consider the distribution of the size of animals in nature, considered within-species. The height of humans follows (almost) a Normal Distribution but it is hard to find mechanism of random walk behind it (this is an observation imparted to the author by Yaneer Bar Yam). Third, uncertainty and opacity lead to power laws, when a statistical mechanism has an error rate which in turn has an error rate, and thus, recursively (Taleb, 2011, 2013). Our approach here is to assume that random variables, under absence of contraints, become power law-distributed. This is the default in the absence of boundedness or compactness. Then, the response, that is, a funtion of the random variable, considered in turn as an “inherited” random variable, will have different properties. If the response is bounded, then the dampening of the tails of the inherited distribution will lead it to bear the properties of the Gaussian, or the class of distributions possessing finite moments of all orders.

expressed as an N summed and scaled standard sigmoid functions: N S HxL ª ‚ k =1 N ak 1 + expH-bk x + ck L + KL (1) where ak . 0L -0. K>2 ). c1 . c1 . The sigmoid shows benefits increasing rapidly (the convex phase). and ¶∂2 SN ¶∂ x2 ¥ 0 for x œ (K>2 . b sets the variance. c1 L. 15N 2 1 S1Hx. 1.. special cases: S 2 Hx . 1. c1 L. 1. a1 . 15L S2Hx. b1 . unlike the common sigmoid.5 S2Hx. ¶). or KR . -2. . 1.KL . 2. ¶∂2 SN ¶∂ x2 < 0 for x œ (K2 . KR ] be a continuous function possessing derivatives HSN LHnL HxL of all orders. The shapes at different calibrations are shown in Figure 1. 1. 1. a2 .. c2 L. ¥ KN . b1 .2 Dose Response. 1. and the standard sigmoid S1 Hx. a1 . ck are norming constants œ R. a2 . the asymptotic response can be lower than the maximum. 1.1. 1. 2. The convex part with positive first detivative has been designated as “antifragile” a sets the height. b1 .5 -1. 25L Fragile Zone -5 5 10 15 20 Harm Slowing Down S2Jx. a1 . or the slope of the sigmoid.nb The Dose Response Let SN HxL: R Ø [KL .0 Antifragile 0. b1 =1 and c1 =0. bk . though in a convex manner. a1 . 1. then increasing at a slower and slower rate until saturation. Assume KL = 0. 1. c2 L. with K1 > K2 ¥ K3 . 2. b2 . Our more general case starts by increasing. S 1 Hx . The Generalized Response Curve. b1 . 1.0 Figure 0. satisfying: i) SN (-¶) =KL ii) SN (¶) =KL + KR N where KR = ‚ak i=1 and (equivalently for the first and last of the following conditions) iii) ¶∂2 SN ¶∂ x2 ¥ 0 for x œ (-¶. As we can see.. with a1 =1. b2 . but the reponse can be actually negative beyond the saturation phase. in which we combined different values of N=2 S2 Hx. -2. and c sets the displacement . K1 ) . as our curves are not monotonically increasing. Harm slows down and becomes “flat” when something is totally broken.

3988 8. : .2.N = 106 ) And the Kurtosis of the inherited distributions drops at higher s thanks to the boundedness of the payoff.3 is infinite. but in-sample will be extremely high. Distribution f.77458 4. . i. finite. and support on the real line.1. ¶). BHa. of course.2.1. Histogram for the different inherited probability distributions (simulations . without loss of generality) as a Student T Distribution with scale s and exponent a. N = 1 .3 HxL S2 H1.0L 1 2 Kurtosis 86. E where B is the Euler Beta function. with power laws at large negative and positive values.20523 Analytical Derivation: Case of the standard sigmoid. Kurtosis for f.a HxL ª a s BA .. and density fs.tLb-1 dt.1. . making the truncation to the left and the right visible. > Figure 0. expressed (for clarity.2. bL ‡ GHaL GHbL ê GHa + bL ‡ Ÿ0 ta-1 H1 .15L S H1.Dose Response.2.nb 3 Properties of the Inherited Probability Distribution Now let x be a random variable with distributed according to a general fat tailed distribution.-1ê2.-2.a HxL: 1+a a a+ x2 s2 2 (2) a 2 1 2 fs.2.08643 4. So we use it as a benchmark to see the drop from the calibration of the response curves.1. Its domain D f = (-¶. but.e.1. 1 The silumation effect of the convex-concave transformations of the terminal probability distribution is shown in Figure 2.15L S H1.

0L gHx.5 1.1. The different inherited probability distributions. . Kurtosis 16 14 12 Out[61]= 10 8 6 4 s 0. 0. 2.3. KR ).4.0 x Figure 0.2 0. 0. 1L gJx. The Kurtosis of the standard drops along with the scale s of the power law .8 1.1. 2. 0.4 0. which can be shown to have a scaled domain Dg = (KL . 2.x +c1 2 (3) a+ g HxL = b12 s2 a b1 s x BJ .0 1. 0. 1. 1.4 Dose Response. 1N 2 3 0.5 2. 1. 0L gHx.1.1. N Ha1.6 0. 2. 1.xL 2 2 a 1 gHxL gHx.0 Figure 0. 2. 1. 1.nb SHxL ª a1 1+expH-b1 x+c1 L g(x) is the inherited distribution. It becomes a+1 2 a1 log a x a1.

in which their version of the utility concept with respect to changes in wealth is concave on the left.Dose Response.nb 5 Remark 1: The inherited distribution from S(x) will have a compact support regardless of the probability distribution of x. and convex on the right). This approach to the dose-response curve is quite general. hence bounded. and can be used outside biology (say in the Kahneman-Tversky prospect theory. Acknowledgments . Conclusion and Remarks We showed the dose-response as the neglected origin of the thin-tailedness of observed distributions in nature.